Newest papers

EpiSEM-Bench: Evaluating Safety and Evidence Recovery in LLM Agents under Missing Evidence

Qijin Zhang, Hanzhen Wang, Yuehan Zhao, Haoyu Dai, Yao Shu, Zhongxiang Dai, Shuang Qiu · 2026 · arXiv

DOI: 10.48550/arXiv.2610.31650 · open-access PDF

Spanish

Cuando una respuesta de una herramienta omite un hecho relevante para la seguridad, un agente puede realizar una acción prohibida a pesar de que la solicitud del usuario es benigna y cada declaración devuelta es verdadera. Introducimos \bench{}, un benchmark ejecutable para evaluar la ejecución insegura, los límites de seguridad y la recuperación de evidencias en los agentes de LLM bajo evidencias faltantes. Contiene 300 pares de tareas inseguras/seguras en seis dominios.

Machine translation (Meta M2M100); the English original follows.

Abstract

Tool-using language model agents rely on environmental facts to decide whether an action is permitted. When a tool response omits a safety-relevant fact, an agent may carry out a prohibited action even though the user request is benign and every returned statement is true. We introduce \bench{}, an executable benchmark for evaluating unsafe execution, safety boundaries, and evidence recovery in LLM agents under missing evidence. It contains 300 matched unsafe/safe task pairs across six domains. By systematically suppressing critical facts, the benchmark measures unsafe execution under critical

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange